Fix stale block content after SQLite INSERT OR REPLACE

[marcobambini]

With SQLite's default recursive_triggers=OFF, INSERT OR REPLACE skips the old row's delete trigger. Deleting a row also keeps its block values, so a row written again as a whole rewrites its block column from the first position and anything past the new value's end stays behind. Replacing block text AAA\nBBB\nCCC with ZZZ produced ZZZ\nBBB\nCCC on replicas and on a later local materialization.

local_block_update now reads the column's stored blocks on that path too. They stay out of the diff — a whole-row write is not an edit of the previous value — and whatever the new value does not rewrite is tombstoned and its value removed. Write failures propagate, so the enclosing statement rolls back rather than committing half a column.

PostgreSQL shares that code and has a milder form of the same problem: the leftovers carry no metadata, so replicas were never affected, but they stayed in the blocks table for the life of the row. They are now retired there too.

Rebased onto main

This PR was rebased after #46, #64, #65, #66 and #67 merged, and now carries a single commit. main had since moved the SQLite insert path onto the shared local_block_insert/local_block_update, so the original SQLite-only dbsync_replace_blocks no longer applied. The fix is ported into the shared block write instead, which keeps one implementation for both backends and preserves the error reporting main added.

Validation

• SQLite: the original 120-cycle replacement regression (recursive triggers off and on; shorter, longer, empty, NULL, duplicate-line and trailing-delimiter values; fractional positions; duplicate delivery; injected write failure) passes unchanged, as does the whole unit suite, including the delete/re-insert test that a naive fix breaks.
• PostgreSQL: new test/postgresql/64_block_rewrite_leftovers.sql — a row recreated with fewer blocks, and an upsert of the whole row, must leave one stored block and deliver the shorter value to a peer. Full suite with ON_ERROR_STOP=on: 557 checks pass on standalone PostgreSQL 17, 555 on Supabase.
• Negative controls against main's code: the SQLite regression fails (replace: recursive=0 round=0 replica=0 mismatch) and the PostgreSQL test fails (shorter rewrite left 3 stored block(s)).

CHANGELOG.md has an [Unreleased] entry; docs/block-replace-fix.md describes the behaviour and the tests.